def forward(self, context_idx, context_mask, context_segements,
query_ids, query_mask, query_segements):
query_emb = self.embed_question(query_ids, query_segements)
context_emb = self.embed_context(context_idx, context_segements)
embedding = torch.cat((query_emb[:, 0], context_emb[:, 0]), dim=1)
start_embedding = self.start_pointer(embedding)
end_embeddings = self.end_pointer(embedding)
start_embedding = mask(start_embedding, context_mask, -1)
end_embeddings = mask(end_embeddings, context_mask, -1)
return start_embedding, end_embeddings
def pointer(self, embeddings, input_mask):
""""""
# size: batch_size, seq_length, 2
embeddings = self.pointer_linear(embeddings)
embeddings = mask(embeddings, input_mask, -2)
# start_embeddings = embeddings[:, :, 0].squeeze(dim=-1)
# end_embeddings = embeddings[:, :, 1].squeeze(dim=-1)
return embeddings
def _RmQ_l(self, l, i, j):
if i == j:
return i
if i > j:
i, j = j, i
w = mask(l[j], i)
return 0 if w == 0 else lsb(w)
def query_pointer(self, embeddings, input_mask): """""" # size: batch_size, seq_length, 2 batch_size, len, dim = embeddings.shape embeddings = self.query_pointor_linear(embeddings) embeddings = reshape_tensor(embeddings, (batch_size, 512, 2)) embeddings = mask(embeddings, input_mask, -2) start_embeddings = embeddings[:, :, 0].squeeze(dim=-1) end_embeddings = embeddings[:, :, 1].squeeze(dim=-1) return start_embeddings, end_embeddings